Verified against Claude · 2026-07-09
Make Claude critique and revise its own draft before you see it
A two-pass prompt that forces Claude to generate a first draft, critique it against a named rubric as if reviewing someone else's work, and then produce a revised version, so you see the improved output instead of doing the editing yourself.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Do this in two explicit passes. Do not skip straight to a polished answer. TASK A 150-word product description for a stainless steel water bottle, for an ecommerce listing. PASS 1 — DRAFT Write a first draft. Do not over-optimize it; a normal-effort first attempt is fine and expected here. PASS 2 — CRITIQUE Now critique that draft as if you are a skeptical editor reviewing someone else's work, not the author defending it. Check it specifically against: - Does it lead with a concrete benefit, not a generic feature list? - Is there a specific, checkable claim (capacity, material, insulation duration) rather than vague superlatives? - Would a skeptical shopper believe this, or does it read like every other listing? Name the weakest part of the draft even if the rest is strong — do not write a critique that is secretly just praise with one soft caveat attached. PASS 3 — REVISION Rewrite the draft to fix everything named in Pass 2. If a criticism from Pass 2 does not get addressed in the revision, say so explicitly and explain why, rather than quietly dropping it. SHOW ALL THREE PASSES Do not show me only the final revision — I want to see the draft, the critique, and the revision, in that order, so I can check the revision actually earned the improvement.
Customize the highlighted detailsoptional — the prompt above already works
Why this works
A single-pass 'write this well' prompt gives Claude no way to catch its own weak choices, because the first plausible draft is also the only draft — there is nothing to compare it against. Explicitly structuring three separate passes forces the model out of author mode and into reviewer mode for the middle step, which is a meaningfully different task: critiquing text you already committed to as good triggers a different, more adversarial read than generating it fresh, similar to why writers benefit from a cooling-off period before self-editing. Naming a specific rubric for the critique pass rather than asking for open-ended feedback prevents the vague-praise failure mode where a model asked to 'critique' its own work produces a soft, hedged paragraph that reads more like reassurance than an actual review — a checkable list of concrete criteria makes 'this is fine as-is' much harder to default to. Requiring that any unaddressed criticism from Pass 2 be explicitly acknowledged in Pass 3, rather than silently dropped, closes the last gap in self-revision loops: without this, a model can perform the appearance of critique and revision while the final output quietly ignores the one finding that was actually hardest to fix.
What you get back
Draft: a generic 150-word description leading with material and dimensions. Critique: "Weakest part — opens with capacity and material, which is what every competing listing does; there is no specific, checkable claim about insulation duration, and 'premium quality' is an unsupported superlative." Revision: description restructured to lead with "keeps drinks cold for 24 hours" as the opening line, "premium quality" replaced with the specific double-wall vacuum insulation spec.
Verified against
Claude Sonnet 4.6 · 2026-07-09
Changelog
- 2026-07-09 — Initial publish, verified against Claude Sonnet 4.6.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
